An intracranial pressure monitoring and early warning system based on imaging features

Abstract

The application relates to the technical field of medical health early warning, in particular to an intracranial pressure monitoring and early warning system based on imaging features, which comprises the following modules: a brain structure twin construction module, which is used for acquiring multi-modal image data and constructing brain structure twins; a brain imaging feature extraction module, which is used for obtaining brain imaging features; an intracranial pressure inversion atlas generation module, which is used for acquiring the distribution response relationship of the brain imaging features and generating an intracranial pressure inversion atlas; an intracranial blood flow regulation capacity quantification module, which is used for acquiring transcranial Doppler ultrasound data and quantifying an intracranial blood flow regulation capacity index; an intracranial pressure state evolution module, which is used for obtaining an intracranial pressure state evolution curve; and an intracranial pressure monitoring and early warning module, which is used for generating graded early warning thresholds and triggering monitoring and early warning. The application realizes dynamic quantitative early warning suitable for individual physiological characteristics by acquiring the evolution trend of intracranial pressure.

Description

Technical Field

[0001] This invention relates to the field of medical and health early warning technology, and in particular to an intracranial pressure monitoring and early warning system based on radiomics features. Background Technology

[0002] Currently, invasive measurement methods remain the gold standard for intracranial pressure monitoring in clinical practice, primarily including external ventricular drainage and intraparenchymal implantation of probes. External ventricular drainage, which involves placing a catheter in the ventricle to drain cerebrospinal fluid and measuring pressure with an external pressure sensor, serves both diagnostic and therapeutic functions and is considered the gold standard for intracranial pressure monitoring. Intraparenchymal implantation of miniature sensors enables continuous monitoring. However, with advancements in medical technology, non-invasive intracranial pressure monitoring methods are also continuously developing, achieving qualitative or semi-quantitative assessments of intracranial pressure by evaluating changes in intracranial structures, cerebral hemodynamic parameters, or ocular physiological indicators.

[0003] Despite the high accuracy of invasive intracranial pressure monitoring, its clinical application has significant limitations. First, invasive procedures carry a considerable risk of complications, including intracranial infection, brain injury, and hemorrhage, and the equipment is expensive. Furthermore, existing non-invasive monitoring methods often employ isolated assessments using single modal indicators, failing to effectively integrate multidimensional information from structural imaging and functional monitoring. This makes it difficult to accurately identify the critical state transitioning from compensated to decompensated stages, thus hindering early warning.

[0004] Therefore, there is an urgent need for an intelligent intracranial pressure monitoring and early warning system that can integrate multimodal data and achieve structural-functional collaborative assessment. Summary of the Invention

[0005] To address the shortcomings of existing technologies and meet the needs of practical applications, this invention provides an intracranial pressure monitoring and early warning system based on radiomics features, comprising the following modules: a cranial-brain structural twin construction module, used to acquire multimodal image data and construct a cranial-brain structural twin based on the multimodal image data; a cranial-brain radiomics feature extraction module, used to extract features from the cranial-brain structural twin to obtain cranial-brain radiomics features; an intracranial pressure inversion map generation module, used to acquire the distribution response relationship of the cranial-brain radiomics features and generate an intracranial pressure inversion map based on the distribution response relationship; an intracranial blood flow regulation capacity quantification module, used to acquire transcranial Doppler ultrasound data and extract dynamic features of cerebral blood flow to quantify the intracranial blood flow regulation capacity index; an intracranial pressure status evolution module, used to fuse the intracranial pressure inversion map and the intracranial blood flow regulation capacity index to obtain an intracranial pressure status evolution curve; and an intracranial pressure monitoring and early warning module, used to dynamically generate graded early warning thresholds based on the intracranial pressure status evolution curves and trigger monitoring and early warning when the intracranial pressure amplitude reaches the graded early warning threshold. This invention improves the timeliness and reliability of intracranial pressure abnormality monitoring, adapts to the actual needs of clinical intensive care, and reduces the risk of missed diagnosis and misdiagnosis.

[0006] Optionally, acquiring multimodal image data and constructing a cranial twin based on the multimodal image data includes the following steps: acquiring computed tomography (CT) images and magnetic resonance imaging (MRI) images as the multimodal image data, and rigidly registering the multimodal image data to obtain multimodal registered images; based on the multimodal registered images, performing layered fine segmentation of the cranial structure using an edge-constrained segmentation network to obtain cranial tissue regions; assigning corresponding biomechanical property parameters to different cranial tissue regions, including elastic modulus, Poisson's ratio, and density; and combining the biomechanical property parameters, using a tetrahedral meshing method to perform three-dimensional meshing of the cranial tissue regions to generate the cranial twin. This invention restores the real anatomical structure and biomechanical characteristics of the cranium, conforms to human physiological reality, and ensures the accuracy and simulation stability of the twin.

[0007] Optionally, the feature extraction from the cranial structural twin to obtain cranial radiomics features includes the following steps: based on the cranial structural twin, simulating the pulsating patterns of cranial tissue driven by the cardiac cycle to obtain the micro-displacement trajectories of key anatomical landmarks; based on the cranial structural twin, fitting the intracranial venous return change characteristics driven by the respiratory cycle to quantify the micro-fluctuation of intracranial volume; combining the micro-displacement trajectories and the micro-fluctuation of intracranial volume to obtain the micro-dynamic feature set of the key anatomical landmarks; and performing feature reconstruction processing on the micro-dynamic feature set to obtain the cranial radiomics features. This invention enhances the correlation between features and intracranial pressure changes, providing highly discriminative and sensitive core feature data for subsequent pressure inversion.

[0008] Optionally, the step of fitting the intracranial venous return variation characteristics driven by the respiratory cycle based on the cranial structural twin to quantify the micro-fluctuation of intracranial volume includes the following steps: the intracranial venous return variation characteristics include peak return, trough return, and return change rate; the volume response of the cranial structural twin is obtained based on the intracranial venous return variation characteristics, and time-domain analysis is performed on the volume response to extract the micro-fluctuation of intracranial volume. This invention enhances the physiological relevance of radiomics features and improves the accuracy of subsequent intracranial pressure inversion.

[0009] Optionally, obtaining the distribution response relationship of the cranial radiomics features and generating an intracranial pressure inversion map based on the distribution response relationship includes the following steps: applying intracranial pressure loads of different gradients to the cranial structural twin and simulating the cranial tissue deformation response using a finite element method; extracting feature values ​​of the cranial radiomics features based on the cranial tissue deformation response and constructing the distribution response relationship based on the feature values; and continuously fitting the distribution response relationship using the response surface methodology to generate the intracranial pressure inversion map. This invention ensures that the map closely matches the actual pressure change pattern, eliminates errors from discrete data, and improves the stability and accuracy of non-invasive intracranial pressure measurement.

[0010] Optionally, acquiring transcranial Doppler ultrasound data and extracting dynamic features of cerebral blood flow to quantify the intracranial blood flow regulation capacity index includes the following steps: extracting the dynamic features of cerebral blood flow from the transcranial Doppler ultrasound data, including peak systolic velocity, end-diastolic velocity, and pulsatility index; acquiring non-invasive blood pressure monitoring data; calculating a hemodynamic transfer function based on the dynamic features of cerebral blood flow and the non-invasive blood pressure monitoring data; and quantifying the intracranial blood flow regulation capacity index based on the hemodynamic transfer function. This invention supplements the intracranial pressure status assessment dimension from the perspective of blood flow regulation, making the overall assessment more comprehensive and more closely aligned with pathophysiological mechanisms.

[0011] Optionally, quantifying the intracranial blood flow regulation capacity index based on the hemodynamic transfer function includes the following steps: extracting frequency domain features, including gain spectrum, phase spectrum, and coherence function, from the hemodynamic transfer function; inputting the frequency domain features into a nonlinear physiological model, and identifying the intracranial blood flow regulation capacity index through model parameters. This invention accurately captures the nonlinear physiological characteristics of blood flow regulation, improves the sensitivity and specificity of intracranial blood flow regulation capacity assessment, and provides reliable data support for intracranial pressure status fusion.

[0012] Optionally, the step of fusing the intracranial pressure inversion map and the intracranial blood flow regulation index to obtain the intracranial pressure status evolution curve includes the following steps: using the cranial radiomics features as input to the intracranial pressure inversion map, and outputting an equivalent value of intracranial pressure after inversion; calculating the dynamic ratio of the equivalent value of intracranial pressure to the amplitude fluctuation of intracranial volume over a continuous time period to obtain the intracranial cavity compliance status; and temporally fusing the intracranial cavity compliance status and the intracranial blood flow regulation index to generate the intracranial pressure status evolution curve. This invention achieves deep temporal fusion of multi-dimensional physiological data, comprehensively reflecting the intracranial cavity compensatory capacity and pressure evolution patterns, and improving the foresight of intracranial pressure status assessment.

[0013] Optionally, the step of dynamically generating a graded warning threshold based on the intracranial pressure status evolution curve and triggering a monitoring warning when the intracranial pressure amplitude reaches the graded warning threshold includes the following steps: performing change feature analysis on the intracranial pressure status evolution curve to obtain a curvature sequence and a change rate sequence; identifying curvature abrupt change points based on the curvature sequence and calculating the average change rate and instantaneous change rate within different time windows based on the change rate sequence; dynamically generating the graded warning threshold based on the curvature abrupt change points, the average change rate, and the instantaneous change rate, combined with historical baseline data; comparing the intracranial pressure amplitude of the intracranial pressure status evolution curve with the graded warning threshold, and triggering a graded warning signal and pushing it to the monitoring terminal when the graded warning threshold is reached. This invention effectively shortens the clinical response time, helps prevent critical complications caused by a sudden increase in intracranial pressure, and improves the level of intelligence in clinical monitoring.

[0014] Optionally, the step of dynamically generating the graded early warning threshold based on the curvature abrupt change point, the average rate of change, and the instantaneous rate of change, combined with historical baseline data, includes the following steps: constructing a curve feature change vector based on the curvature abrupt change point, the average rate of change, and the instantaneous rate of change, and performing similarity matching with the historical baseline data to obtain a matching result; adjusting the basic early warning threshold according to the matching result, and generating the graded early warning threshold according to the early warning risk level. This invention improves the reliability of early warning results, reduces interference from invalid alarms, and adapts to the individualized monitoring needs of patients. Attached Figure Description

[0015] Figure 1 This is a framework diagram of an intracranial pressure monitoring and early warning system based on radiomics features according to an embodiment of the present invention; Figure labeling: A1, Cranial-Brain Structure Twin Construction Module; A2, Cranial-Brain Imaging Feature Extraction Module; A3, Intracranial Pressure Inversion Atlas Generation Module; A4, Intracranial Blood Flow Regulation Capacity Quantification Module; A5, Intracranial Pressure Status Evolution Module; A6, Intracranial Pressure Monitoring and Early Warning Module. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 An embodiment of the present invention provides an intracranial pressure monitoring and early warning system based on radiomics features, comprising the following modules: a brain structural twin construction module A1, a brain radiomics feature extraction module A2, an intracranial pressure inversion map generation module A3, an intracranial blood flow regulation capacity quantification module A4, an intracranial pressure status evolution module A5, and an intracranial pressure monitoring and early warning module A6.

[0019] In this embodiment, the cranial twin construction module A1 is used to acquire multimodal image data and construct a cranial twin based on the multimodal image data; including the following steps: S11. Acquire computed tomography (CT) images and magnetic resonance imaging (MRI) images as the multimodal image data, and perform rigid registration on the multimodal image data to obtain multimodal registered images.

[0020] In this embodiment, multimodal imaging data of the patient is acquired, including computed tomography (CT) images and magnetic resonance imaging (MRI) images. CT images are obtained using a standard head scanning protocol with a slice thickness of 1 mm and a matrix of 512×512, acquiring density information of the entire skull and brain tissue. MRI images are obtained using a T1-weighted three-dimensional gradient echo sequence with a slice thickness of 1 mm and a matrix of 256×256, acquiring high-resolution structural information of the brain tissue.

[0021] Furthermore, to eliminate spatial positional deviations caused by differences in patient positioning and scanning parameters, rigid registration was performed on the two types of image data. The MRI image was used as the reference image (due to its superior contrast for soft tissues), while the CT image was used as the floating image. A registration optimization algorithm based on Normalized Mutual Information (NMI) was employed, achieving global alignment through affine transformations (including translation, rotation, and scaling). During registration, an initial iteration step size of 0.1 was set, and a multi-resolution strategy (from coarse to fine) was used to accelerate convergence: coarse registration was first performed on the downsampled low-resolution image, followed by fine registration on the original resolution image. The NMI value was calculated after each iteration, and the process terminated when the rate of change was less than 0.1% or the maximum number of iterations (200) was reached. After registration, the CT image was resampled to the spatial resolution of the MRI image (1mm × 1mm × 1mm) using bilinear interpolation, resulting in a spatially fully aligned multimodal registered image, providing accurate anatomical positional correspondences for subsequent segmentation.

[0022] S12. Based on the multimodal registration image, the cranial structure is finely segmented by an edge-constrained segmentation network to obtain the cranial tissue region.

[0023] In this embodiment, a deep learning segmentation network is used to perform hierarchical and fine segmentation of the cranial structure based on multimodal registered images.

[0024] Specifically, the deep learning segmentation network is based on the U-Net architecture. The encoder part contains 5 downsampling layers, each using a 3×3×3 convolutional kernel with a stride of 2 and the number of channels being 32, 64, 128, 256, and 512 respectively. The decoder part contains 5 upsampling layers, each using transposed convolutions, with the number of channels being 256, 128, 64, 32, and 16 respectively.

[0025] Furthermore, to improve segmentation accuracy, an edge constraint module is introduced into the decoder of the network structure. This module extracts gradient information from the image using the Sobel operator (Sobel for short) to generate an edge attention map, and multiplies it element-wise with the feature map of the intermediate layer of the network to enhance the edge response to the cranial structure.

[0026] Specifically, the network input consists of registered CT and MRI dual-channel images, and the output is a multi-class segmentation probability map, with categories including: skull, dura mater, arachnoid mater, gray matter, white matter, lateral ventricle, third ventricle, fourth ventricle, great cerebral vein, and circle of basilar artery.

[0027] In this embodiment, the training data comes from public datasets and a hospital's self-built database, and the annotation was completed by domain experts. The loss function is a weighted combination of Dice loss and boundary loss, with weights of 0.8 and 0.2, respectively, to simultaneously optimize region overlap and boundary accuracy. The network is trained using an Adaptive Moment Estimation (Adam) optimizer with an initial learning rate of 0.001, a batch size of 4, and 200 iterations.

[0028] Furthermore, the output multi-class segmentation probability map is post-processed with Conditional Random Field (CRF) to eliminate isolated pixel noise, and finally obtains the three-dimensional binary mask of each tissue region, namely the cranial tissue region mask, to obtain the cranial tissue region.

[0029] S13. For different brain tissue regions, corresponding biomechanical property parameters are assigned, including elastic modulus, Poisson's ratio, and density.

[0030] In this embodiment, based on the brain tissue region mask obtained from hierarchical fine segmentation, each voxel or mesh node is assigned corresponding biomechanical property parameters, including elastic modulus ( Poisson's ratio ) and density ( ).

[0031] Specifically, the parameters are taken from clinical experience values ​​and are set differently for different tissue types: the skull is regarded as an isotropic linear elastic material, with an elastic modulus of 15 GPa, Poisson's ratio of 0.3, and density of 1.5 g / cm³. 3 The dura mater is considered a linear elastic membrane structure with an elastic modulus of 31.5 MPa, a Poisson's ratio of 0.45, and a density of 1.13 g / cm³. 3 The arachnoid membrane is considered an incompressible material, with an elastic modulus of 10 kPa, a Poisson's ratio of 0.49, and a density of 1.0 g / cm³. 3The gray and white matter of the brain are considered viscoelastic materials and described using the generalized Maxwell model (GMM). Their instantaneous elastic moduli are taken as 2 kPa and 1.5 kPa, respectively, with a relaxation time constant of 0.2 s, a Poisson's ratio of 0.49, and a density of 1.04 g / cm³. 3 The ventricles are considered cavities filled with cerebrospinal fluid, giving them fluid properties, with a density of 1.0 g / cm³. 3 Blood vessels are considered as isotropic linear elastic materials with an elastic modulus of 1 MPa, a Poisson's ratio of 0.4, and a density of 1.06 g / cm³. 3 .

[0032] It should be noted that parameter assignment is implemented through a script, automatically mapping based on the label values ​​of the segmentation mask. To ensure the stability of the finite element calculation, spatial smoothing is performed on the parameters to avoid numerical oscillations caused by abrupt parameter changes at adjacent tissue boundaries. After assignment, a voxel model containing attribute information is generated, providing a physical property basis for subsequent meshing.

[0033] S14. Combining the biomechanical property parameters, the cranial tissue region is three-dimensionally meshed using a tetrahedral meshing method to generate the cranial structure twin.

[0034] In this embodiment, based on the voxel model, the tetrahedral meshing method is used to perform three-dimensional meshing of the cranial tissue region to generate a finite element mesh.

[0035] First, a three-dimensional region growing algorithm is used to extract the surface contours of each tissue region from the voxel model to generate a surface mesh.

[0036] Then, tetrahedral filling is performed using open-source mesh generation tools or commercial software. The maximum mesh cell size is set to 2 mm to accommodate the geometric features of intracranial microstructures (such as vascular branches); local refinement is performed at tissue interfaces (such as skull-dura mater, gray matter-white matter boundaries) with a cell size of 1 mm to improve the calculation accuracy of contact areas.

[0037] It should be noted that the mesh quality check includes indicators such as minimum dihedral angle (>20°) and Jacobian ratio (>0.3). Unqualified cells are subject to local re-optimization.

[0038] Ultimately, the generated tetrahedral mesh contains approximately 3 to 5 million cells, each assigned a corresponding attribute (determined by the tissue type at its location). The mesh node coordinates are aligned with the original image space, forming a twin of the cranial structure with geometric and physical consistency.

[0039] The cranial-brain structural twin not only accurately reproduces the anatomical morphology of the patient's cranium, but also integrates biomechanical properties, providing a high-fidelity foundation for subsequent microdynamic simulations (such as cardiac cycle pulsation and respiratory cycle venous return changes).

[0040] It should be noted that the cranial structural twin can be periodically updated based on newly acquired multimodal imaging data. When the patient's condition changes (such as the occurrence of cerebral edema, ventricular enlargement, or progression of space-occupying lesions), the system can update the geometric morphology and biomechanical property parameters of the cranial structural twin to ensure that the subsequent intracranial pressure inversion map matches the current pathological state. The update cycle is set by the clinician based on the stability of the patient's condition, with a typical value of once every 24 to 72 hours, or triggered when the doctor determines that the condition has changed significantly.

[0041] In this embodiment, the cranial radiomics feature extraction module A2 is used to extract features from the cranial structural twin to obtain cranial radiomics features; including the following steps: S21. Based on the aforementioned cranial twin structure, simulate the pulsating pattern of cranial tissue driven by the cardiac cycle to obtain the micro-displacement trajectory of key anatomical landmarks.

[0042] In this embodiment, based on a cranial twin, the pulsation pattern of cranial tissue driven by the cardiac cycle is simulated through explicit finite element dynamics analysis.

[0043] First, the patient's electrocardiogram signal is acquired simultaneously, and the cardiac cycle waveform is extracted from it. The pressure changes during ventricular systole and diastole are converted into time-domain load curves acting on the blood vessel wall.

[0044] Specifically, the peak systolic pressure is defined as 120 mmHg and the trough diastolic pressure is defined as 80 mmHg. The time-domain pressure waveform is decomposed into the fundamental frequency (approximately 1 Hz) and its harmonic components using a fast Fourier transform.

[0045] Subsequently, the aforementioned pressure load was applied to the walls of the main arteries in the craniosynostosis, while the outer surface of the skull was set as a fixed constraint to simulate the rigid boundary conditions of the skull. An explicit dynamic solver was used for time-domain solution with a time step of 0.001 s, and the total calculation time covered three complete cardiac cycles to ensure stable response.

[0046] It should be noted that during the solution process, the focus is on the three-dimensional displacement response of key anatomical landmarks such as the pineal gland, falx cerebri, and tentorium cerebelli. As a midline structure, the pineal gland is sensitive to changes in intracranial pressure; the falx cerebri and tentorium cerebelli, as folds of the dura mater, can reflect the overall pulsatile conduction of brain tissue through their displacement.

[0047] Finally, after the calculations are completed, the x, y, and z displacement values ​​of each key anatomical landmark at each time step are extracted to form a displacement-time series. This series is then bandpass filtered to remove low-frequency respiratory interference and equipment noise, yielding the micro-amplitude displacement trajectory driven by the cardiac cycle.

[0048] S22. Based on the aforementioned cranial twin, fit the characteristics of intracranial venous return changes driven by the respiratory cycle to quantify the minute fluctuations in intracranial volume.

[0049] Within the respiratory cycle, the respiratory cycle waveform is extracted based on synchronously acquired respiratory signals. The respiratory cycle is divided into the inspiratory phase, the expiratory phase, and the apnea phase. During the inspiratory phase, the negative pressure in the thoracic cavity increases, promoting venous return; during the expiratory phase, the negative pressure in the thoracic cavity decreases, slowing venous return; and the apnea phase is a smooth transition period.

[0050] In this embodiment, the respiratory phase changes are converted into time-varying pressure boundary conditions acting on the superior vena cava and internal jugular vein: the venous outlet pressure is set to 5 mmHg during the inspiratory phase and 10 mmHg during the expiratory phase, while linear interpolation is maintained during the apnea phase. This pressure load is applied to the main veins in the cranial twin to simulate the dynamic changes in venous return driven by respiration.

[0051] Furthermore, a fluid-structure interaction finite element method (FSB) was used for calculations, where blood flow was described by the Navier-Stokes equations, and the vessel walls and brain tissue were constitutively linearly elastic. The solution yielded a time-varying venous return flow curve, from which three key features were extracted: the peak flow (maximum flow rate during inspiration), the trough flow (minimum flow rate during expiration), and the rate of change of flow (the rate of change in flow between the peak and trough, i.e., the change in flow rate per unit time). These features collectively constitute the characteristics of intracranial venous return changes, comprehensively characterizing the modulation amplitude and dynamic response speed of venous return during the respiratory cycle.

[0052] In this embodiment, in order to quantify the minute fluctuation of intracranial volume, the intracranial venous return change characteristics are used as input loads and reapplied to the venous vessels of the cranial-brain structural twin. A fluid-structure interaction solver is used to calculate the ventricular and brain tissue volume response caused by the venous return change.

[0053] Specifically, the flow rates corresponding to the peak and trough of venous return are used as boundary conditions to obtain the maximum and minimum filling states of venous blood volume during the cardiac cycle; the rate of change of venous return is used as a dynamic excitation to evaluate the response speed of volume changes.

[0054] Furthermore, the volume change curves of the ventricular system and brain tissue at each respiratory phase are obtained as the volume response. Time-domain analysis is performed on these curves: the maximum and minimum volumes within each respiratory cycle are extracted, and the difference between them is calculated as the volume fluctuation amplitude; the first derivative of the volume change curve is extracted to obtain the volume change rate; the average value of the fluctuation amplitude over multiple respiratory cycles (usually 10 cycles) is taken as the final quantified intracranial volume fluctuation.

[0055] It should be noted that the slight fluctuation in intracranial volume physiologically includes the combined contribution of fluctuations in blood volume and cerebrospinal fluid volume. This is because the respiratory cycle regulates the total intracranial blood volume by affecting venous return, while changes in thoracic pressure also affect the drainage of cerebrospinal fluid into the subarachnoid space of the spinal cord.

[0056] S23. Combine the micro-displacement trajectory and the micro-fluctuation of the intracranial volume to obtain the micro-dynamic feature set of the key anatomical landmarks.

[0057] First, the micro-displacement trajectories of key anatomical landmarks are parameterized. For each landmark, the following parameters are calculated from the displacement-time series: displacement amplitude (average peak displacement over three cardiac cycles), displacement velocity (first derivative of the displacement series, maximum value), displacement acceleration (second derivative of the displacement series, maximum value), and displacement phase delay (delay time relative to the peak value of the ECG R wave).

[0058] Simultaneously, the volume fluctuation amplitude, volume change rate, and volume response phase (delay time relative to the trough of the respiratory signal) are extracted from the volume response driven by the respiratory cycle.

[0059] Subsequently, the above parameters were spatiotemporally aligned and fused: using the cardiac cycle as the time reference, the volume fluctuation parameters were aligned to the same time axis through interpolation; using key anatomical landmarks as the spatial reference, the volume fluctuation parameters were mapped to the locations of each landmark.

[0060] Specifically, for each key anatomical landmark, a multidimensional feature vector is constructed, including: x-axis displacement amplitude, y-axis displacement amplitude, z-axis displacement amplitude, composite displacement amplitude, displacement velocity, displacement acceleration, displacement phase delay, local volume fluctuation amplitude, local volume change rate, and volume response phase. This multidimensional feature vector represents the microdynamic characteristics of that landmark.

[0061] Finally, the microdynamic features of all key anatomical landmarks were compiled to form a microdynamic feature set. This feature set not only includes the conduction characteristics of cardiac pulsation (through displacement trajectory) but also the modulation characteristics of local volume by the respiratory cycle (through volume fluctuation), realizing the information fusion of the cardiac-respiratory dual physiological cycles.

[0062] S24. Perform feature reconstruction processing on the microdynamic feature set to obtain the cranial imaging omics features.

[0063] In this embodiment, the microdynamic feature set characterizes the local dynamic response at key anatomical landmarks, but its sparse distribution characteristics make it difficult to directly match the full-field mapping requirements of the subsequent intracranial pressure inversion atlas. Therefore, feature reconstruction processing is used to convert discrete microdynamic features into spatially continuous brain radiomics features.

[0064] First, a feature matrix representation is constructed: the microdynamic feature set is constructed as a feature matrix, including key anatomical landmarks and their multidimensional feature vectors. This feature matrix retains the spatial coordinate information of each landmark.

[0065] Secondly, spatial continuity mapping is performed: considering that the tissue response caused by changes in intracranial pressure is spatially continuous, radial basis function interpolation is used to extend the sparse marker features to the whole brain spatial region.

[0066] Specifically, for each feature dimension, the feature values ​​of all marker points in that dimension are used as interpolation node values, and the spatial coordinates of the marker points are used as node positions. A radial basis function interpolation model is constructed, satisfying the following relationship: in, For the first Each feature dimension in spatial location Interpolation estimate at point, Let be the coordinate vector of the target space point. This represents the total number of key anatomical landmarks. An index for key anatomical landmarks. For the first The feature dimension corresponds to the _th _th The radial basis function weight coefficients for each marker point For radial basis functions, For the first The spatial coordinate vectors of the marker points For the first Transpose of the column vector of polynomial coefficients under each feature dimension For the first The polynomial constant term under each feature dimension.

[0067] It should be noted that the coefficients of the polynomial terms are determined by solving the system of linear equations.

[0068] By substituting the coordinates of each voxel in the whole-brain 3D mesh into the interpolation model, the estimated values ​​of each voxel coordinate in each feature dimension are obtained, thereby generating multiple 3D feature maps, the spatial resolution of each feature map being perfectly aligned with the original image.

[0069] Then, feature fusion and dimensionality reduction were performed: the 3D feature maps generated by the radial basis function interpolation model have feature redundancy, and the contribution of some feature dimensions to intracranial pressure inversion varies. Principal component analysis was used to reduce the dimensionality of the 3D feature maps: each feature map was flattened into a one-dimensional vector, the covariance matrix was calculated and eigenvalues ​​were extracted, and principal components with a cumulative contribution rate of over 95% were retained. The high-dimensional feature maps were projected onto the low-dimensional principal component space to obtain the fused feature map, which both compressed the feature dimensions and preserved the core dynamic information.

[0070] Finally, feature map standardization and format encapsulation are performed: to ensure the reconstructed feature maps are compatible with the input requirements of subsequent intracranial pressure inversion maps, each fused feature map is z-score normalized to a mean of 0 and a standard deviation of 1, eliminating dimensional differences. The normalized feature maps are then stacked along the channel dimension to form a multi-channel three-dimensional feature tensor, which serves as the feature for cranial radiomics.

[0071] In this embodiment, the intracranial pressure inversion map generation module A3 is used to obtain the distribution response relationship of the cranial radiomics features and generate an intracranial pressure inversion map based on the distribution response relationship; including the following steps: S31. Apply intracranial pressure loads of different gradients to the cranial twin structure and obtain the cranial tissue deformation response through simulation using a finite element solver.

[0072] In this embodiment, a series of discrete intracranial pressure load gradients are established based on a craniosynaptic twin. According to clinical practice, to cover the complete range from normal physiological to pathological states, the load gradients are set as follows: a range of 5 mmHg-50 mmHg with intervals of 5 mmHg, resulting in a total of 10 gradients. For each gradient, the pressure load is applied to the surface of the brain tissue and the ventricular wall to simulate the mechanical effect of intracranial pressure on the brain parenchyma. The outer surface of the skull is set as a fixed constraint to simulate the rigid boundary conditions of the cranial cavity; a contact pair is established between the brain tissue and the skull, allowing relative sliding but not intrusion.

[0073] Steady-state statics analysis was performed using a finite element method (FEM). A viscoelastic constitutive model was adopted for the brain tissue, with the following material parameters: instantaneous shear modulus of 2 kPa, long-term shear modulus of 1 kPa, and relaxation time constant of 0.2 s. During the solution process, the convergence of the model was monitored to ensure that the displacement residual at each load step was less than 10. -6 After calculation, the full-field displacement distribution under each load gradient is extracted to form a three-dimensional deformation field. This deformation field contains the x, y, and z displacement values ​​of each node, with a resolution consistent with the original twin (approximately 1 mm). 3 (Volumetrics).

[0074] In this embodiment, to verify the rationality of the simulation, the displacement values ​​of 10 anatomical landmarks (such as the pineal gland, the anterior horn of the lateral ventricle, and the genu of the corpus callosum) were randomly selected to ensure that they exhibit monotonically nonlinear changes with increasing load, and that the displacement amplitude is within a physiologically reasonable range (e.g., the displacement of the pineal gland is about 0.5mm-2.0mm under a load of 50mmHg). Finally, the pressure-deformation field is output as the deformation response of the cranial tissue, laying the data foundation for feature extraction.

[0075] S32. Based on the cranial tissue deformation response, extract the feature values ​​of the cranial radiomics features, and construct the distribution response relationship based on the feature values.

[0076] In this embodiment, for the simulated cranial tissue deformation response under each pressure gradient, feature values ​​corresponding to cranial radiomics features are extracted.

[0077] First, for the cranial tissue deformation response obtained under each pressure gradient, the deformation response is quantitatively characterized in the form of displacement field, strain field and stress field of three-dimensional spatial grid nodes, and a one-to-one spatial mapping relationship is established with the cranial radiomics characteristics.

[0078] Specifically, the region of interest (including gray matter, white matter, cerebrospinal fluid, and ventricles) obtained from cranial image segmentation is used as the benchmark for spatial registration. The deformation response results are back-mapped to the original anatomical coordinate system of the region of interest through a nonlinear registration algorithm to ensure that each voxel position corresponds to a set of deformation response parameters.

[0079] Secondly, for the pre-screened and determined brain radiomics features, including morphological features, texture features, and mechanical features, quantitative extraction was performed from the brain tissue deformation response. The extraction process adopted a voxel-based feature calculation method: for morphological features, the local volume change rate was obtained by calculating the Jacobian determinant of each voxel before and after deformation, and the anisotropy index was obtained based on the deformation gradient tensor decomposition; for texture features, a second-order tensor field was constructed based on the deformation displacement field and its eigenvalues ​​were calculated, thereby obtaining texture parameters characterizing tissue slippage and fiber stretching; for mechanical features, the equivalent mechanical response values ​​of each voxel were directly output from the stress-strain relationship based on the constitutive model.

[0080] It should be noted that all feature values ​​are stored in the form of three-dimensional feature maps, and each feature map has the same spatial resolution and coordinate system as the original image.

[0081] Finally, the various feature values ​​extracted under each pressure gradient were organized into feature vector matrices. Principal component analysis or canonical correlation analysis was used to reduce the dimensionality of the features and fuse them to eliminate redundant information. The pressure gradient value was used as the independent variable and the fused feature vector was used as the dependent variable. A continuous pressure-feature response curve was fitted using a generalized additive model or Gaussian process regression to form a distribution response relationship that characterizes the radiomic features of cranial tissue under different pressure loads with changes in spatial location and gradient.

[0082] The distributed response relationship is solidified in the form of a parametric model, which includes both the global variation law of eigenvalues ​​with pressure gradient and the local heterogeneity of eigenresponse in each region of interest, thus providing a calculable and traceable quantitative basis for the subsequent accurate assessment of cranial mechanical state.

[0083] S33. The distribution response relationship is continuously fitted using the response surface methodology to generate the intracranial pressure inversion map.

[0084] In this embodiment, intracranial pressure is used as the independent variable based on the response surface methodology. The eigenvalues ​​of each cranial radiomics feature were used as the dependent variable. Construct quadratic polynomial response surface models respectively, satisfying the following relationship: in, For eigenvalues, For undetermined coefficients, Intracranial pressure, This is the error term.

[0085] It should be noted that for features exhibiting an S-shaped growth trend (such as pineal gland displacement), the logistic function (Logistic for short) can be used for fitting; for features exhibiting exponential growth (such as ventricular compression rate in high-pressure segments), exponential or power functions can be used for fitting.

[0086] Furthermore, the least squares method was used for model fitting to determine the values ​​of the undetermined coefficients. After fitting, residual analysis was performed on each model to ensure that the residuals were randomly distributed and without a clear pattern; the goodness of fit was tested, requiring that the coefficient of determination of each eigenvalue be greater than 0.95.

[0087] In this embodiment, after completing the response surface fitting for all feature values, the fitted functions are integrated into a unified intracranial pressure inversion map. This map is mathematically expressed as: given any intracranial pressure value... The corresponding eigenvalues ​​can be calculated using each response surface model.

[0088] Furthermore, the atlas is encapsulated as a callable function or lookup table: when the measured cranial radiomics feature vector is input, the corresponding intracranial pressure value can be solved by reverse lookup using an optimization algorithm, thus realizing the reverse mapping from feature value to pressure.

[0089] In this embodiment, the intracranial blood flow regulation capacity quantification module A4 is used to acquire transcranial Doppler ultrasound data and extract dynamic features of cerebral blood flow to quantify the intracranial blood flow regulation capacity index; including the following steps: S41. Extract the dynamic features of cerebral blood flow from the transcranial Doppler ultrasound data, including peak systolic velocity, end-diastolic velocity, and pulsatility index.

[0090] In this embodiment, a 2MHz pulse wave probe was used to collect transcranial Doppler (TCD) data of the middle cerebral artery through the temporal window. The sampling frequency was set to 128Hz and the sampling depth was set to 50mm-60mm to obtain a clear blood flow velocity waveform.

[0091] It should be noted that during the collection process, the patient needs to maintain calm breathing, avoid swallowing and head shaking, and record at least 10 complete cardiac cycles continuously.

[0092] Furthermore, after the original TCD signal is bandpass filtered to remove baseline drift and high-frequency noise, the start and end points of a single cardiac cycle are identified by an automatic thresholding method.

[0093] Specifically, using the lowest end-diastolic flow velocity as the periodic division point, the continuous waveform is divided into several independent cardiac cycles. First, for each cardiac cycle, a peak detection algorithm is used to locate the point of maximum systolic flow velocity, and the velocity value at this point is recorded as the systolic peak flow velocity. Next, the lowest flow velocity point at the end of diastole (i.e., before the start of the next cardiac cycle) is located, and the velocity value at this point is recorded as the end-diastolic flow velocity. Finally, the pulsatility index is calculated by combining the systolic peak flow velocity and the end-diastolic flow velocity, satisfying the following relationship: in, The pulsatility index, Peak velocity during contraction. For the end-diastolic flow rate, The average flow velocity is denoted as .

[0094] It should be noted that the average flow rate is obtained by integrating and averaging the flow rate waveform over the entire cardiac cycle.

[0095] In this embodiment, to ensure the stability of feature extraction, the average values ​​of PSV, EDV, and PI for 10 consecutive cardiac cycles are taken as the dynamic features of cerebral blood flow during the current acquisition period.

[0096] In an optional embodiment, the above feature extraction process can be repeated in clinical practice, for example, by collecting data every 5 minutes to form a time series of dynamic features of cerebral blood flow, providing basic data for subsequent hemodynamic analysis. All feature values ​​are stored in numerical form and accompanied by a collection timestamp to be time-aligned with synchronously acquired non-invasive blood pressure data.

[0097] S42. Obtain non-invasive blood pressure monitoring data, and calculate the hemodynamic transfer function based on the cerebral blood flow dynamic characteristics and the non-invasive blood pressure monitoring data.

[0098] In this embodiment, a non-invasive continuous blood pressure monitoring device is used to continuously acquire arterial blood pressure waveforms from the patient's fingertips or wrists. The sampling frequency is consistent with that of the TCD data to ensure the time synchronization of the two signals. The blood pressure signal and the TCD flow velocity signal are time-aligned, and the peak value of the ECG R wave is used as a common time reference point to eliminate phase deviation caused by device delay.

[0099] After alignment, the blood pressure and flow velocity data were divided into several 30-second time windows, with adjacent windows overlapping by 50% to improve the stability of frequency estimation. For each time window, Welch's Averaged Periodogram Method (Welch) was used to estimate the power spectral density and cross-spectral density: the data within the window was segmented (each segment was 2 seconds long, i.e., 256 sampling points), and a Hanning window was applied to each segment. The Fourier transform of each segment was calculated, and the amplitude of the Fourier transform results of each segment was squared and averaged to obtain the power spectral density of that window. The conjugate product of the Fourier transform results of the two signals (blood pressure and flow velocity) within the same window was calculated and averaged to obtain the cross-spectral density of that window.

[0100] In this embodiment, the hemodynamic transfer function is defined as the ratio of the blood pressure-flow velocity cross-spectral density to the blood pressure self-power spectral density.

[0101] S43. Quantify the intracranial blood flow regulation capacity index based on the hemodynamic transfer function.

[0102] In the frequency domain, the hemodynamic transfer function can be decomposed into a gain spectrum (amplitude response) and a phase spectrum (phase response). The gain spectrum reflects the amplification factor of blood pressure changes to flow velocity changes, while the phase spectrum reflects the lag time of the flow velocity response relative to blood pressure changes. Furthermore, by calculating the ratio of the square of the cross-spectral density amplitude to the blood pressure and flow velocity spectra, a coherence function is obtained. This coherence function is used to assess the linear correlation between blood pressure and flow velocity at a specific frequency; the closer the coherence function value is to 1, the more reliable the transfer function estimate at that frequency.

[0103] In this embodiment, for the transfer function calculated for each time window, the mean gain, mean phase, and peak coherence function in the extremely low frequency band (0.02Hz-0.07Hz) are extracted.

[0104] Specifically, the gain spectrum is integrated and averaged over the frequency range to obtain the average gain of that frequency band, which reflects the overall efficiency of cerebral blood flow autoregulation. The lower the gain, the less blood pressure fluctuations are transmitted to cerebral blood flow, and the stronger the autoregulation capability. The phase mean is obtained by calculating the weighted average of the phase spectrum in that frequency band. The closer the phase is to 0 (i.e., the flow velocity response is synchronized with blood pressure changes), the faster the regulation response. The peak value of the coherence function is extracted from the maximum coherence value in that frequency band to evaluate the reliability of the estimated transfer function. Only when the peak value of the coherence function is greater than 0.5 are the gain and phase data of this window included in subsequent analysis.

[0105] Furthermore, by using a sliding time window process, the time series of frequency domain feature parameters under each window are obtained and used as input for the nonlinear physiological model.

[0106] In this embodiment, a nonlinear physiological model based on the Laguerre-Volterra Network (LVN) is employed. This model can describe the nonlinear dynamic relationship between arterial blood pressure and cerebral blood flow velocity. The model structure is as follows: taking arterial blood pressure as input and cerebral blood flow velocity as output, the input signal is expanded through a set of orthogonal Laguerre basis functions, and the system response is represented as a second-order Volterra series of the outputs of each basis function.

[0107] The parameters to be identified in the model include linear kernel coefficients, nonlinear kernel coefficients, and coefficients of delay terms of each order. The least squares algorithm is used for parameter identification to prevent overfitting.

[0108] During the identification process, frequency domain features were used as initial parameter constraints for the model. The objective function was optimized using gradient descent, with 200 iterations and an adaptive Adam optimizer for the learning rate. After model convergence, quantitative indicators reflecting the autoregulation ability of cerebral blood flow were extracted from the identified model parameters: the zeroth moment of the linear kernel was used as the basic regulation coefficient, and the second moment of the nonlinear kernel was used as the regulation nonlinearity. The two were weighted and fused to obtain the intracranial blood flow regulation ability index.

[0109] It should be noted that a higher intracranial blood flow regulation capacity index indicates a stronger cerebrovascular regulation function, which can more effectively buffer the impact of blood pressure fluctuations on cerebral blood flow. This index is time-series aligned with the synchronously calculated intracranial pressure inversion map to provide basic data for the subsequent intracranial pressure situation evolution module, which integrates structure and function.

[0110] In this embodiment, the intracranial pressure status evolution module A5 is used to fuse the intracranial pressure inversion map and the intracranial blood flow regulation capacity index to obtain the intracranial pressure status evolution curve; including the following steps: S51. Using the cranial imaging features as input to the intracranial pressure inversion map, the equivalent value of intracranial pressure is output after inversion.

[0111] In this embodiment, the brain radiomics features are input into the intracranial pressure inversion map to output the equivalent value of intracranial pressure. Since the measured feature values ​​may not completely match the feature values ​​of any discrete pressure point, the equivalent value of intracranial pressure is obtained by solving an optimization algorithm.

[0112] Specifically, the Newton-Raphson iterative method is used for inversion: the objective function is defined as the mean square error between the measured feature vector and the predicted feature vector of the inverted map. The initial value for iteration is set to 15 mmHg (normal mean intracranial pressure), and the iteration step size is adaptively adjusted: the step size is increased when the objective function decreases rapidly and decreased when it decreases slowly. The convergence condition is set to the change in the objective function between two consecutive iterations being less than 10. -6 Or it can reach the maximum number of iterations, 50.

[0113] Furthermore, after obtaining the equivalent value of intracranial pressure, confidence interval estimation is required. Based on the statistical distribution of the fitting residuals of the inversion map, the 95% confidence interval of the equivalent value of intracranial pressure is calculated to evaluate the reliability of the inversion results. If the confidence interval width is less than 5 mmHg, the inversion results are considered reliable; if it is greater than 5 mmHg, it indicates that the current feature set has high noise or is in the boundary region of the inversion map, and smoothing processing is required in conjunction with subsequent time-series information.

[0114] In this embodiment, the final output intracranial pressure equivalent value is stored in numerical form, along with the acquisition timestamp and confidence level marker, for use in subsequent dynamic ratio calculation. This inversion process can be executed in real time after acquiring new brain radiomics features each time, forming a time series of intracranial pressure equivalent values.

[0115] S52. Calculate the dynamic ratio of the equivalent value of intracranial pressure to the slight fluctuation of intracranial volume over a continuous time period to obtain the intracranial compliance status.

[0116] In this embodiment, the amplitude fluctuation of intracranial volume and the equivalent value of intracranial pressure are obtained, both based on the same time axis (the sampling interval is set according to clinical needs). The time axis is divided into continuous sliding time windows, with adjacent windows overlapping by 50% to ensure the continuity of situational evolution.

[0117] Within each time window, the dynamic ratio of the equivalent value of intracranial pressure to the slight fluctuation of intracranial volume was calculated. The relationship between the equivalent value of intracranial pressure (dependent variable) and the slight fluctuation of intracranial volume (independent variable) within the window was fitted using a linear regression method, and the regression coefficient is the dynamic ratio within that window.

[0118] The reciprocal of the dynamic ratio represents intracranial compliance (i.e., the volume change caused by a unit change in pressure). Physiologically, during the compensatory phase, compliance is high (dynamic ratio is low) due to the buffering effect of cerebrospinal fluid and venous blood; as the compensatory mechanism depletes, compliance decreases (dynamic ratio increases). In addition to calculating the mean of the dynamic ratio, it is also necessary to calculate the coefficient of variation (the ratio of standard deviation to mean) of the dynamic ratio within the window, reflecting the stability of compliance; and the rate of change of the dynamic ratio over time (first derivative), reflecting the trend of compliance change. The mean, coefficient of variation, and rate of change over time are integrated into a three-dimensional vector as the intracranial compliance status of the current window.

[0119] For example, a mean less than 0.1 and a small coefficient of variation indicate good compliance; a mean greater than 0.3 and a positive rate of change over time indicate rapid deterioration of compliance. Repeating the above calculations for each sliding window yields a time series of cranial compliance status, reflecting the dynamic process of the evolution of cranial compensatory capacity over time.

[0120] In an optional embodiment, to further improve the robustness of the situation assessment, the output compliant situation sequence can be subjected to Kalman filtering to eliminate short-term noise fluctuations and retain the true physiological change trends.

[0121] S53. The intracranial compliance status and the intracranial blood flow regulation capacity index are fused in a time sequence to generate the intracranial pressure status evolution curve.

[0122] In this embodiment, the two sequences are time-aligned and standardized (z-score normalization) to eliminate dimensional differences; a weighted temporal fusion method is used to construct a fusion feature vector that satisfies the following relationship: in, To fuse feature vectors, For dynamic weights, The normalized cranial cavity compliance posture score is the overall score. This is the normalized index of intracranial blood flow regulation capacity.

[0123] It should be noted that the intracranial compliance posture comprehensive score is a one-dimensional score obtained by dimensionality reduction of the three-dimensional features through principal component analysis; furthermore, the determination of dynamic weights is based on clinical priors: in the early stages of compensation, compliance posture contributes significantly ( During the critical period of decompensation, blood flow regulation capacity is a better indicator of decompensation risk; therefore, increasing blood flow regulation at this time is recommended. ( The weights are adaptively calculated over time based on the trend of data changes within the sliding window.

[0124] In this embodiment, the fused feature vector is post-processed: a filter is used to smooth it and eliminate high-frequency fluctuations; the fused feature vector is mapped to the 0-100 range, where 0 represents stable intracranial pressure and good compensation, and 100 represents the critical state of decompensation; key event points are marked on the curve, such as the time point when the dynamic ratio in the compliance state exceeds 0.3 and the time point when the blood flow regulation capacity index is lower than 30. The final intracranial pressure state evolution curve is displayed on the monitoring terminal.

[0125] In this embodiment, the intracranial pressure monitoring and early warning module A6 is used to dynamically generate graded early warning thresholds based on the intracranial pressure status evolution curve, and to trigger a monitoring and early warning when the intracranial pressure amplitude reaches the graded early warning threshold; including the following steps: S61. Analyze the change characteristics of the intracranial pressure status evolution curve to obtain the curvature sequence and the rate of change sequence.

[0126] In this embodiment, the first and second derivatives of the intracranial pressure situation evolution curve are calculated. The first derivative is calculated using the central difference method, reflecting the instantaneous rate of change in the situation evolution; the second derivative is calculated using the difference of the first derivative; the curvature is calculated using the first and second derivatives, and the larger the curvature value, the more severe the curvature of the curve at that point.

[0127] To ensure numerical stability, when the first derivative is too large, the curvature is normalized and mapped to the 0-1 interval. After calculation, a curvature sequence and a rate of change sequence of the same length as the original curve are obtained. These sequences characterize the evolution curve of intracranial pressure status from two dimensions: geometric shape (curvature) and dynamic change (rate). The curvature sequence is used to identify the turning point from stable to rapid change (i.e., the compensatory critical point), and the rate of change sequence is used to quantify the speed of status evolution.

[0128] S62. Identify curvature abrupt change points based on the curvature sequence, and calculate the average rate of change and instantaneous rate of change within different time windows based on the rate of change sequence.

[0129] First, for the curvature sequence, a sliding window thresholding method is used to identify curvature abrupt changes. The window length is set to 30 seconds (i.e., 30 data points), and the mean curvature within the window is calculated. and standard deviation When the curvature value at the center point of the window exceeds When this point is reached, it is marked as a candidate mutation point.

[0130] In this embodiment, to avoid misjudgments caused by noise, a two-way verification method is used: taking the candidate mutation point as a boundary, 10 data points are extended forward and backward respectively, and the difference in the mean curvature of the two segments is calculated. If the difference is greater than 0.3, the point is confirmed as a curvature mutation point. The curvature mutation point corresponds to the critical position where the intracranial pressure status evolution curve shifts from a flat area to a steep increase area, physiologically representing a critical moment when the intracranial compensation mechanism begins to deplete and is about to enter a decompensated state. For the identified mutation points, their occurrence time and curvature value are recorded as an important basis for subsequent threshold adjustment.

[0131] Secondly, for the rate of change sequence, the average rate of change and instantaneous rate of change within different time windows are calculated. The instantaneous rate of change is directly calculated using the first derivative sequence, i.e., the instantaneous rate of change at each time point. The average rate of change is calculated using multiple time scales: three time scales are set: short window (60 seconds), medium window (300 seconds), and long window (600 seconds). For each window, the arithmetic mean of the rates of change within the window is calculated to obtain the short-term, medium-term, and long-term average rates of change. The short-term average rate reflects the intensity of recent trend fluctuations, the medium-term average rate reflects the persistence of the trend, and the long-term average rate reflects the direction of baseline evolution.

[0132] Meanwhile, to eliminate the influence of extreme values, outliers with a rate of change exceeding the mean ± 3 standard deviations within the window are removed when calculating the average rate.

[0133] After all calculations are completed, the following characteristic parameters are output: a set of curvature abrupt change points (including timestamps and curvature values), an instantaneous rate of change sequence (per second), and three sets of average rate of change sequences (60-second, 300-second, and 600-second windows).

[0134] S63. Based on the curvature abrupt change point, the average rate of change, and the instantaneous rate of change, combined with historical baseline data, the graded early warning threshold is dynamically generated.

[0135] In this embodiment, the set of curvature abrupt change points, the instantaneous rate of change sequence, and the average rate of change sequence are integrated into the curve feature change vector.

[0136] Meanwhile, historical baseline data, which are the mean and standard deviation of the corresponding characteristic parameters of patients during their previous stable periods (such as when there are no symptoms of increased intracranial pressure and no abnormalities on imaging), are stored in the database.

[0137] Furthermore, similarity matching uses Mahalanobis distance calculation to obtain matching results. Mahalanobis distance ( This can eliminate the influence of differences in the dimensions and correlations of various feature dimensions. The smaller the distance, the closer the current situation is to a stable state; the larger the distance, the more serious the deviation from normal. Based on clinical experience, the matching results are graded as follows: when... At that time, the matching result was "normal"; At that time, it was "attention"; At that time, it was a "warning"; The situation is "critical". This matching result provides a quantitative basis for subsequent threshold adjustments.

[0138] In this embodiment, the basic warning threshold is adjusted based on the matching results, and graded warning thresholds are generated according to the warning risk level. The basic warning thresholds are determined based on clinical guidelines or historical data statistics: 60 points for low risk, 75 points for medium risk, and 85 points for high risk.

[0139] The adjustment rules are dynamically adjusted based on the matching results and the current rate of change: When the matching result is "normal", the basic threshold remains unchanged; when the matching result is "attention", each threshold is lowered by 5 points (i.e., 55 / 70 / 80) to reduce the warning trigger threshold and increase sensitivity; when the matching result is "alert", fine adjustments are made according to the instantaneous rate of change: if it is greater than 0.5 (rapid deterioration of the situation), the threshold is lowered by 10 points (50 / 65 / 75), and if it is less than or equal to 0.5, it is lowered by 5 points (55 / 70 / 80); when the matching result is "critical", regardless of the rate of change, the threshold is directly lowered by 15 points (45 / 60 / 70), and a high-risk warning is forcibly triggered.

[0140] Furthermore, curvature abrupt change points are used as an independent constraint: if the time difference between the current moment and the most recent curvature abrupt change point is less than 60 seconds, regardless of the matching result, the medium-risk and high-risk thresholds are lowered by an additional 5 points to enhance the response to compensatory critical points. After adjustment, the final graded early warning thresholds are generated: low-risk threshold... Medium-risk threshold and high-risk threshold .

[0141] S64. Compare the intracranial pressure amplitude of the intracranial pressure status evolution curve with the graded warning threshold. When the graded warning threshold is reached, a graded warning signal is triggered and pushed to the monitoring terminal.

[0142] In this embodiment, the intracranial pressure amplitude and graded warning threshold of the intracranial pressure status evolution curve are continuously acquired. The comparison rule uses a duration-based condition to avoid false alarms triggered by single-point noise. A duration threshold is set: a low-risk warning requires more than 5 consecutive seconds. A medium-risk warning needs to last for more than 3 consecutive seconds. High-risk warnings must be issued continuously for more than 1 second. When the corresponding conditions are met, the corresponding graded warning signal is triggered. The graded warning signals adopt graded coding: Level 1 warning (low risk) is a yellow flashing signal, Level 2 warning (medium risk) is an orange flashing signal accompanied by an intermittent buzzer sound, and Level 3 warning (high risk) is a red flashing signal accompanied by a continuous buzzer sound.

[0143] Upon triggering an alert, the system automatically records the current situation score, trigger threshold, curvature change point information, and matching results, generating an alert event log.

[0144] Furthermore, the tiered early warning signal push adopts a multi-channel parallel strategy: First, the early warning information is sent to the central monitoring station through the hospital information system and prominently displayed on the large screen at the nurses' station; second, it is pushed to the mobile terminals of attending physicians and nurses via wireless network, and the information content includes patient information, early warning level, current intracranial pressure status score and suggested intervention measures; finally, for high-risk early warnings, the system automatically triggers the audible and visual alarm of the bedside monitor and calls the on-duty doctor.

[0145] In an optional embodiment, to prevent warning fatigue, the system sets the minimum interval between warnings of the same level to 10 minutes, meaning that warnings of the same level will not be pushed again within 10 minutes after a warning is triggered. However, higher-level warnings can break this restriction. Warning records are stored in a cloud database for subsequent evaluation and model optimization.

[0146] The system provided by this invention realizes a complete closed loop from data acquisition to intelligent early warning, providing real-time, accurate, and personalized reference for clinical intracranial pressure monitoring and early warning.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An intracranial pressure monitoring and early warning system based on radiomics features, characterized in that, Includes the following modules: A cranial-brain structural twin construction module is used to acquire multimodal image data and construct a cranial-brain structural twin based on the multimodal image data; The cranial radiomics feature extraction module is used to extract features from the cranial structural twin to obtain cranial radiomics features. The intracranial pressure inversion map generation module is used to obtain the distribution response relationship of the brain radiomics features and generate an intracranial pressure inversion map based on the distribution response relationship. The intracranial blood flow regulation capacity quantification module is used to acquire transcranial Doppler ultrasound data and extract dynamic features of cerebral blood flow to quantify the intracranial blood flow regulation capacity index. The intracranial pressure situation evolution module is used to fuse the intracranial pressure inversion map and the intracranial blood flow regulation capacity index to obtain the intracranial pressure situation evolution curve; The intracranial pressure monitoring and early warning module is used to dynamically generate graded early warning thresholds based on the intracranial pressure status evolution curve, and to trigger monitoring and early warning when the intracranial pressure amplitude reaches the graded early warning threshold.

2. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 1, characterized in that, The acquisition of multimodal image data and the construction of a cranial twin based on the multimodal image data include the following steps: Acquire computed tomography (CT) images and magnetic resonance imaging (MRI) images as the multimodal image data, and perform rigid registration on the multimodal image data to obtain multimodal registered images; Based on the multimodal registered images, the cranial structure is finely segmented into layers using an edge-constrained segmentation network to obtain cranial tissue regions. For different regions of the cranial tissue, corresponding biomechanical property parameters are assigned, including elastic modulus, Poisson's ratio, and density. Based on the biomechanical property parameters, the cranial tissue region is meshed in three dimensions using a tetrahedral meshing method to generate a cranial structural twin.

3. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 1, characterized in that, The process of extracting features from the cranial twin to obtain cranial radiomics features includes the following steps: Based on the aforementioned cranial twin, the pulsating pattern of cranial tissue driven by the cardiac cycle is simulated to obtain the micro-displacement trajectory of key anatomical landmarks. Based on the aforementioned cranial structural twin, the characteristics of intracranial venous return changes driven by the respiratory cycle were fitted to quantify the minute fluctuations in intracranial volume. The micro-dynamic feature set of the key anatomical landmarks is obtained by combining the micro-displacement trajectory and the micro-fluctuation of the intracranial volume; The brain radiomics features are obtained by performing feature reconstruction processing on the microdynamic feature set.

4. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 3, characterized in that, The step of fitting the intracranial venous return characteristics driven by the respiratory cycle based on the cranial structural twin to quantify the minute fluctuations in intracranial volume includes the following steps: The characteristics of intracranial venous return include peak return, trough return, and rate of change in return. The volume response of the cranial twin is obtained based on the intracranial venous return change characteristics, and time-domain analysis is performed on the volume response to extract the intracranial volume fluctuation.

5. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 1, characterized in that, The process of acquiring the distribution response relationship of the cranial radiomics features and generating an intracranial pressure inversion map based on the distribution response relationship includes the following steps: Different gradients of intracranial pressure loads were applied to the cranial twin, and the cranial tissue deformation response was obtained by simulation using a finite element solver. Based on the cranial tissue deformation response, feature values ​​of the cranial radiomics features are extracted, and the distribution response relationship is constructed based on the feature values. The response surface methodology is used to continuously fit the distributed response relationship to generate the intracranial pressure inversion map.

6. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 1, characterized in that, The acquisition of transcranial Doppler ultrasound data and extraction of dynamic features of cerebral blood flow to quantify the intracranial blood flow regulation capacity index includes the following steps: The dynamic features of cerebral blood flow, including peak systolic velocity, end-diastolic velocity, and pulsatility index, are extracted from the transcranial Doppler ultrasound data. Acquire non-invasive blood pressure monitoring data, and calculate the hemodynamic transfer function based on the dynamic characteristics of cerebral blood flow and the non-invasive blood pressure monitoring data; The intracranial blood flow regulation capacity index is quantified based on the hemodynamic transfer function.

7. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 6, characterized in that, The quantification of the intracranial blood flow regulation capacity index based on the hemodynamic transfer function includes the following steps: Frequency domain features, including gain spectrum, phase spectrum, and coherence function, are extracted based on the hemodynamic transfer function. The frequency domain features are input into a nonlinear physiological model, and the intracranial blood flow regulation capacity index is obtained by identifying the model parameters.

8. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 3, characterized in that, The process of fusing the intracranial pressure inversion map and the intracranial blood flow regulation capacity index to obtain the intracranial pressure status evolution curve includes the following steps: The brain radiomics features are used as input to the intracranial pressure inversion map, and the equivalent value of intracranial pressure is output after inversion. Over a continuous time period, the dynamic ratio of the equivalent value of intracranial pressure to the slight fluctuation of intracranial volume is calculated to obtain the intracranial compliance status. The intracranial compliance status and the intracranial blood flow regulation capacity index are fused in a time series to generate the intracranial pressure status evolution curve.

9. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 1, characterized in that, The process of dynamically generating graded early warning thresholds based on the intracranial pressure status evolution curve, and triggering monitoring and early warning when the intracranial pressure amplitude reaches the graded early warning threshold, includes the following steps: The changes in the intracranial pressure state evolution curve were analyzed to obtain the curvature sequence and the rate of change sequence. The curvature abrupt change points are identified based on the curvature sequence, and the average rate of change and instantaneous rate of change within different time windows are calculated based on the rate of change sequence. Based on the curvature abrupt change point, the average rate of change, and the instantaneous rate of change, combined with historical baseline data, the graded early warning threshold is dynamically generated; The intracranial pressure amplitude of the intracranial pressure status evolution curve is compared with the graded warning threshold. When the graded warning threshold is reached, a graded warning signal is triggered and pushed to the monitoring terminal.

10. The intracranial pressure monitoring and early warning system based on radiomics features according to claim 9, characterized in that, The process of dynamically generating the graded early warning threshold based on the curvature abrupt change point, the average rate of change, and the instantaneous rate of change, combined with historical baseline data, includes the following steps: Based on the curvature abrupt change point, the average rate of change, and the instantaneous rate of change, a curve feature change vector is constructed, and a similarity matching is performed with the historical baseline data to obtain a matching result; The basic warning threshold is adjusted based on the matching results, and the graded warning threshold is generated according to the warning risk level.

Images